Customizing your knowledge base on Amazon Bedrock for large and complex documents using Amazon Textract
Amazon Web Services Rushabh Lokhande
AWS says raw utility bills were too messy for RAG. Textract cleans them first, so Bedrock can answer more reliably.
Based on reporting by Amazon Web Services, Rushabh Lokhande — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Customer service teams can drown in utility bills. The files arrive in PDFs, DOCX, TXT, HTML, XLSX, and PNG, and the useful bits are buried in tables, scans, and odd layouts. The AWS post starts from a familiar failure: feed those documents straight into a retrieval system, and the model misses details or invents them.
That is the core problem here. A raw RAG setup struggled to pull out account numbers, due dates, payment amounts, and other billing details with enough consistency to trust. So AWS is pushing a preprocessing step with Amazon Textract before the documents ever reach Amazon Bedrock. Textract extracts text from the source files, then the pipeline cleans and enriches the output so the knowledge base has something more reliable to work with.
The workflow is fairly mechanical, which is part of the appeal. A shell script creates the CloudFormation stack. That stack sets up the Lambda execution role, a Lambda layer, two Lambda functions, one S3 bucket, an OpenSearch Serverless cluster, and an Amazon Bedrock knowledge base with its IAM role. After that, users create a raw_files folder in the S3 bucket, upload the sample utility bills, and let the document-parser Lambda kick off Textract jobs.
Processed files land in parsed_files, then a second Lambda converts them into TXT and saves them in parsed_kb_documents. From there, the knowledge base is synced in the Bedrock console and can be queried with Amazon Nova Micro. AWS says the point is to make document-heavy support work faster and more accurate, especially when the alternative is a human combing through a pile of mixed-format bills.
The other notable piece is the cautionary note. AWS calls out Guardrails and grounding validation as production controls to reduce harmful outputs, protect sensitive data, and catch hallucinations when responses aren’t backed by the source documents. That is the right instinct. Document AI is only useful when it stops pretending messy input is clean.
My take — AI-written commentary, not fact-checked reporting
This is the boringly sensible version of enterprise AI, which is usually the best kind. If the documents are messy, clean them before asking a model to be clever. The industry still loves skipping that step and then acting surprised when the chatbot starts freelancing.
Read more about this at: Amazon Web Services